An Efficient Compressive Convolutional Network for Unified Object Detection and Image Compression

Xichuan Zhou, Lang Xu, Shujun Liu, Yingcheng Lin, Lei Zhang, Cheng ZhuoAAAI 2019 · DOI

Abstract

In edge computing environments such as smart surveillance and drones, power and computational resources are severely constrained. Traditional systems handle object detection and image compression as separate, sequential tasks, leading to high latency and hardware costs. This paper introduces the Compressive Convolutional Network (CCN), a framework that unifies these tasks by repurposing the initial layer of an object detection model to serve as a compressive sensing encoder. By integrating a regularization term for incoherence into the training process, the CCN achieves a 3.1x to 5.0x improvement in efficiency over conventional designs while maintaining high detection accuracy and superior image reconstruction quality.

Introduction

Wireless embedded systems face a significant challenge: the "double whammy" of computationally expensive deep learning models for object detection and the high battery consumption required to transmit high-resolution video. To mitigate transmission costs, images are typically compressed using standard methods like JPEG or traditional Compressive Sensing (CS). However, these methods require dedicated processing steps, adding to the system's overall latency and hardware footprint.

Comparison of conventional and unified design for object detection and image compression.

Researchers from Chongqing and Zhejiang Universities proposed the Compressive Convolutional Network (CCN) to address this. As shown in the unified design above, the CCN framework eliminates the need for separate sub-sampling hardware by making the neural network's feature extraction layer perform the compression simultaneously with detection.

Methodology

The CCN is built upon the YOLOv2 architecture and utilizes three primary technical innovations:

  1. Convolution as Embedding: The authors recognized that the first layer of a CNN is mathematically a linear transformation, which can function as a "measurement matrix" for compressive sensing. By applying a uniform pooling layer to the output of this first convolution, they create a compressed representation of the input image.
  2. Incoherence Regularization: To ensure the compressed image can be reconstructed, the measurement matrix must satisfy an "incoherence condition." The authors implemented this as a regularization term ($R_W$) within the loss function during training.
  3. Two-Stage Training: A specialized training strategy was employed to balance detection and compression. The network is first trained with the incoherence penalty to optimize compression kernels, which are then frozen while the rest of the network is fine-tuned for detection accuracy.

For image reconstruction at the receiver (e.g., a cloud server), the system utilizes the standard Orthogonal Matching Pursuit (OMP) algorithm.

Experimental Results

The model was evaluated using the PASCAL VOC and MS COCO datasets for detection, and the Berkeley BSD100 dataset for reconstruction quality.

Efficiency and Accuracy

The CCN-YOLO framework demonstrated a 3.1x to 5.0x increase in efficiency regarding combined latency compared to decoupled systems. Remarkably, this efficiency did not come at the cost of performance. On the VOC dataset, CCN achieved 78.1 mAP, nearly matching the baseline YOLOv2 (78.6 mAP). On the more complex COCO dataset, it reached 43.8 mAP, compared to the original 44.0 mAP.

Compression Performance

In terms of image reconstruction, the CCN significantly outperformed traditional compressive sensing methods. As illustrated in the performance graph below, the CCN consistently maintains a higher Peak Signal-to-Noise Ratio (PSNR) across various compression rates.

PSNR performance of CCN compared to other compressive sensing methods across various compression rates.

Specifically, the reconstruction quality was 3.0 dB to 5.2 dB higher than Gaussian or random convolution matrix approaches. The system is capable of processing images at 40 frames per second on a standard GPU, making it suitable for real-time edge applications.

Conclusion and Future Work

The Compressive Convolutional Network proves that learned, deterministic kernels in a CNN can serve as effective compressive sensing encoders with zero additional front-end computational overhead. While the current back-end reconstruction via OMP is effective, it remains computationally intensive and is best suited for cloud-side processing. Future research will focus on developing a dedicated neural network for the reconstruction phase to further reduce latency and improve the fidelity of the recovered images.

Summary updated

BibTeX
@inproceedings{zhou2019efficient,
  author = {Xichuan Zhou and Lang Xu and Shujun Liu and Yingcheng Lin and Lei Zhang and Cheng Zhuo},
  title = {{An Efficient Compressive Convolutional Network for Unified Object Detection and Image Compression}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33015949},
}
Figure